Evolving Soft Robots For Control Using a Gene Regulatory Network

This repository contains the Python code which accompanies my BSc Computer Science Dissertation, where I investigated how soft robots could be controlled and designed by an evolved Gene Regulatory Network (GRN) to complete locomotion and object-manipulation tasks.

For this research project, I used EvoGym to simulate Voxel-based Soft Robots (VSR). I also designed and implemented a developmental model and a decentralised controller in Python that uses a GRN to both design the robot's architecture and control its movements. In my application, populations of GRNs are evolved with an Evolutionary Algorithm, which I implemented with Python's multiprocessing library to take advantage of performing fitness evaluations in parallel on multiple CPU cores.

Project Structure

This repository was forked from EvoGym so that my classes and methods can be integrated within EvoGym's simulator. My methods can be found under the "gene_regulatory_network" directory. Within this directory are three folders, each representing a different major iteration in the project's development:

• gene_regulatory_network/initial - Contains the initial code for conducting evolutionary experiments using a GRN to optimise the robot's movements with fixed designs.
• gene_regulatory_network/refinement pt.1 - Contains the code for the first refinement iteration, which builds upon the previous implementation to improve the model's results. This contains an updated Evolutionary Algorithm which implements a crossover operator to combine genetic information from parents, and Tournament Selection to select the best candidates. Experiments for this iteration are still concerned with optimising the robot's controller only.
• gene_regulatory_network/refinement pt.2 (co-evolution) - Contains the code for the final iteration of the project, where a different approach is considered to improve the model's performance. This includes an implementation of a developmental model and a decentralised controller, which allows the GRN to be co-evolved to both control the robot's behaviour and specify its design.